[llvm] [TTI] Return a more sensible cost for histogram intrinsic. (PR #97397)

[Graham Hunter via llvm-commits]

================
@@ -508,11 +513,39 @@ static bool isUnpackedVectorVT(EVT VecVT) {
          VecVT.getSizeInBits().getKnownMinValue() < AArch64::SVEBitsPerBlock;
 }
 
+static InstructionCost getHistogramCost(const IntrinsicCostAttributes &ICA) {
+  Type *BucketPtrsTy = ICA.getArgTypes()[0]; // Type of vector of pointers
+  Type *EltTy = ICA.getArgTypes()[1];        // Type of bucket elements
+
+  // Only allow (32b and 64b) integers or pointers for now...
----------------
huntergr-arm wrote:

The histcnt instruction only supports 32b and 64b elements natively. I haven't implemented legalization for smaller types yet, so I'm excluding them for now. Once the loop vectorize changes have landed I intend to implement legalization on AArch64, and will update this function to reflect that.

https://github.com/llvm/llvm-project/pull/97397